Tormenta: geomagnetic Kp nowcast

Tormenta predicts the planetary Kp index (0–9, the standard 3-hourly measure of geomagnetic disturbance) from the solar wind measured upstream of Earth. It is a CT-378 Machine Learning project (NED University, Fall 2026) by Alizeh and HRK. Code: https://github.com/Syeda-Aliza-Ayaz/Tormenta

Model card

Task	Regression: hourly Kp nowcast with an 80% prediction interval
Model	scikit-learn HistGradientBoostingRegressor (point), plus q10/q90 quantile models with a conformal (CQR) margin
Inputs	IMF |B|, Bx, By, Bz (GSM), proton speed, density, temperature, F10.7 solar flux; 66 engineered features (derived physics, lags, trailing means, position in the 3-h Kp block, time of day/year)
Not used as inputs	Any geomagnetic index (Kp, Dst, AE, …), so the nowcast does not wait for magnetometer data
Training data	NASA OMNI2 hourly, 1998 to 2025-07; labels are GFZ Potsdam definitive Kp
Selection	Time-ordered CV in 1998–2016, validation 2017–2019, one final test on 2020 to 2026-01
Live inputs	NOAA SWPC propagated solar wind (1-min, time-shifted to Earth), averaged to hours like OMNI

Test results (2020 to 2026-01, 52,848 hours, includes the May 2024 G5 storm)

Metric	Tormenta	Persistence baseline
RMSE	0.573	0.862
MAE	0.437	0.643
R²	0.808	0.566
RMSE on storm hours (Kp ≥ 5−)	0.846	1.368
80% interval coverage	80.8%	–

Intended use and limitations

• Intended use: education and situational awareness. It is not an operational warning service; for that, use NOAA SWPC or GFZ.
• Big storms are under-predicted: about −0.6 Kp on average for G3+ hours, because such storms are rare in the training data.
• Sudden onsets show up 1–2 h late within a 3-hour Kp block.
• Lead time: at most about 1 h, the travel time of the solar wind from L1 to Earth.
• Manual inputs (predict_steady) are assumed to have held steady for the past 3 days, because the model also uses lags and 24-h averages.

Use the model

Live demo: https://tormenta.streamlit.app/

This repository contains the model (models/current.joblib), the feature and prediction code it needs (src/), and the F10.7 history file used for the trailing solar-flux mean. Requires Python 3.12 and the exact package versions in requirements.txt; the model is a pickled scikit-learn 1.9.1 object.

import sys
from huggingface_hub import snapshot_download

path = snapshot_download("aliza-ayaz/tormenta")
sys.path.insert(0, path)

from src.predict import load_bundle, nowcast, predict_steady

bundle = load_bundle(f"{path}/models/current.joblib")

# Live: fetch NOAA SWPC real-time solar wind, predict every recent hour
pred, hourly = nowcast(bundle)
print(pred.tail(3))            # kp_pred, kp_lo, kp_hi (80% interval), ahead

# Manual: one set of conditions, assumed steady for the past 3 days
print(predict_steady(bundle, {"b_scalar": 20, "bx": 3, "by_gsm": 5, "bz_gsm": -15, "speed": 600,
                              "density": 10, "temperature": 2e5, "f107": 150}))

The bundle is a dict with point, q_lo, q_hi (fitted HistGradientBoostingRegressors), features (the 66 column names), interval_margin, trained_until and version. Build features only with src.features.build_features, which is the same code used in training.

Data and licences

NASA/GSFC Space Physics Data Facility OMNIWeb; GFZ German Research Centre for Geosciences Kp and F10.7 (CC BY 4.0); NOAA Space Weather Prediction Center real-time data (public domain).